pacman::p_load(olsrr, corrplot, ggpubr, sf, spdep,
GWmodel, tmap, tidyverse, gtsummary, kableExtra)In-class Exercise 04
Geographically Weighted Regression
Geographically weighted regression (GWR) is a spatial statistical technique that takes non-stationary variables into consideration (e.g., climate; demographic factors; physical environment characteristics) and models the local relationships between these independent variables and an outcome of interest (aka dependent variable).
This exercise will create a hedonic pricing model to explain the effect of locational factors on condominium prices in Singapore in 2015. A number of independent variables on about the property (e.g. size, type of property) and its surroundings (e.g. proximity to amenities) will be used.
1. Setting Up
Loading Packages
sf and spdep: spatial data handling and spatial weights
tidyverse: manipulation of attribute data and plotting visualisations (aspatial)
tmap: creating map visualisations
corrplot and ggpubr: multivariate data analysis and visualisation
olsrr: building OLS and performing diagnostics tests
GWmodel: calibrating geographical weighted models
gtsummary: create publication-ready analytical and summary tables
Loading Data
The geospatial dataset that will be used is the Singapore Masterplan 2014 subzone planning boundaries.
mpsz <- st_read(dsn="data/geospatial",
layer="MP14_SUBZONE_WEB_PL")Reading layer `MP14_SUBZONE_WEB_PL' from data source
`D:\lins-92\ISSS624\In-class_EX04\data\geospatial' using driver `ESRI Shapefile'
Simple feature collection with 323 features and 15 fields
Geometry type: MULTIPOLYGON
Dimension: XY
Bounding box: xmin: 2667.538 ymin: 15748.72 xmax: 56396.44 ymax: 50256.33
Projected CRS: SVY21When importing geospatial files, we should check the projection:
st_crs(mpsz)Coordinate Reference System:
User input: SVY21
wkt:
PROJCRS["SVY21",
BASEGEOGCRS["SVY21[WGS84]",
DATUM["World Geodetic System 1984",
ELLIPSOID["WGS 84",6378137,298.257223563,
LENGTHUNIT["metre",1]],
ID["EPSG",6326]],
PRIMEM["Greenwich",0,
ANGLEUNIT["Degree",0.0174532925199433]]],
CONVERSION["unnamed",
METHOD["Transverse Mercator",
ID["EPSG",9807]],
PARAMETER["Latitude of natural origin",1.36666666666667,
ANGLEUNIT["Degree",0.0174532925199433],
ID["EPSG",8801]],
PARAMETER["Longitude of natural origin",103.833333333333,
ANGLEUNIT["Degree",0.0174532925199433],
ID["EPSG",8802]],
PARAMETER["Scale factor at natural origin",1,
SCALEUNIT["unity",1],
ID["EPSG",8805]],
PARAMETER["False easting",28001.642,
LENGTHUNIT["metre",1],
ID["EPSG",8806]],
PARAMETER["False northing",38744.572,
LENGTHUNIT["metre",1],
ID["EPSG",8807]]],
CS[Cartesian,2],
AXIS["(E)",east,
ORDER[1],
LENGTHUNIT["metre",1,
ID["EPSG",9001]]],
AXIS["(N)",north,
ORDER[2],
LENGTHUNIT["metre",1,
ID["EPSG",9001]]]]We need to perform a transformation to the EPSG:3414 projection. Although they are both called SVY21 projections, there are some differences in CRS.
mpsz_svy21 <-st_transform(mpsz, 3414)
st_crs(mpsz_svy21)Coordinate Reference System:
User input: EPSG:3414
wkt:
PROJCRS["SVY21 / Singapore TM",
BASEGEOGCRS["SVY21",
DATUM["SVY21",
ELLIPSOID["WGS 84",6378137,298.257223563,
LENGTHUNIT["metre",1]]],
PRIMEM["Greenwich",0,
ANGLEUNIT["degree",0.0174532925199433]],
ID["EPSG",4757]],
CONVERSION["Singapore Transverse Mercator",
METHOD["Transverse Mercator",
ID["EPSG",9807]],
PARAMETER["Latitude of natural origin",1.36666666666667,
ANGLEUNIT["degree",0.0174532925199433],
ID["EPSG",8801]],
PARAMETER["Longitude of natural origin",103.833333333333,
ANGLEUNIT["degree",0.0174532925199433],
ID["EPSG",8802]],
PARAMETER["Scale factor at natural origin",1,
SCALEUNIT["unity",1],
ID["EPSG",8805]],
PARAMETER["False easting",28001.642,
LENGTHUNIT["metre",1],
ID["EPSG",8806]],
PARAMETER["False northing",38744.572,
LENGTHUNIT["metre",1],
ID["EPSG",8807]]],
CS[Cartesian,2],
AXIS["northing (N)",north,
ORDER[1],
LENGTHUNIT["metre",1]],
AXIS["easting (E)",east,
ORDER[2],
LENGTHUNIT["metre",1]],
USAGE[
SCOPE["Cadastre, engineering survey, topographic mapping."],
AREA["Singapore - onshore and offshore."],
BBOX[1.13,103.59,1.47,104.07]],
ID["EPSG",3414]]We can view the geographical extent of the dataset:
st_bbox(mpsz_svy21) xmin ymin xmax ymax
2667.538 15748.721 56396.440 50256.334 Now let’s load the aspatial dataset:
condo_resale = read_csv("data/aspatial/Condo_resale_2015.csv")glimpse(condo_resale)Rows: 1,436
Columns: 23
$ LATITUDE <dbl> 1.287145, 1.328698, 1.313727, 1.308563, 1.321437,…
$ LONGITUDE <dbl> 103.7802, 103.8123, 103.7971, 103.8247, 103.9505,…
$ POSTCODE <dbl> 118635, 288420, 267833, 258380, 467169, 466472, 3…
$ SELLING_PRICE <dbl> 3000000, 3880000, 3325000, 4250000, 1400000, 1320…
$ AREA_SQM <dbl> 309, 290, 248, 127, 145, 139, 218, 141, 165, 168,…
$ AGE <dbl> 30, 32, 33, 7, 28, 22, 24, 24, 27, 31, 17, 22, 6,…
$ PROX_CBD <dbl> 7.941259, 6.609797, 6.898000, 4.038861, 11.783402…
$ PROX_CHILDCARE <dbl> 0.16597932, 0.28027246, 0.42922669, 0.39473543, 0…
$ PROX_ELDERLYCARE <dbl> 2.5198118, 1.9333338, 0.5021395, 1.9910316, 1.121…
$ PROX_URA_GROWTH_AREA <dbl> 6.618741, 7.505109, 6.463887, 4.906512, 6.410632,…
$ PROX_HAWKER_MARKET <dbl> 1.76542207, 0.54507614, 0.37789301, 1.68259969, 0…
$ PROX_KINDERGARTEN <dbl> 0.05835552, 0.61592412, 0.14120309, 0.38200076, 0…
$ PROX_MRT <dbl> 0.5607188, 0.6584461, 0.3053433, 0.6910183, 0.528…
$ PROX_PARK <dbl> 1.1710446, 0.1992269, 0.2779886, 0.9832843, 0.116…
$ PROX_PRIMARY_SCH <dbl> 1.6340256, 0.9747834, 1.4715016, 1.4546324, 0.709…
$ PROX_TOP_PRIMARY_SCH <dbl> 3.3273195, 0.9747834, 1.4715016, 2.3006394, 0.709…
$ PROX_SHOPPING_MALL <dbl> 2.2102717, 2.9374279, 1.2256850, 0.3525671, 1.307…
$ PROX_SUPERMARKET <dbl> 0.9103958, 0.5900617, 0.4135583, 0.4162219, 0.581…
$ PROX_BUS_STOP <dbl> 0.10336166, 0.28673408, 0.28504777, 0.29872340, 0…
$ NO_Of_UNITS <dbl> 18, 20, 27, 30, 30, 31, 32, 32, 32, 32, 34, 34, 3…
$ FAMILY_FRIENDLY <dbl> 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0…
$ FREEHOLD <dbl> 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1…
$ LEASEHOLD_99YR <dbl> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…The condo dataset contains transactions by condominium blocks. It can be transformed into a geospatial dataset using the latitude and longitude data:
condo_resale.sf <- st_as_sf(condo_resale,
coords=c("LONGITUDE", "LATITUDE"),
crs=4326) %>%
st_transform(crs=3414)We can plot the two geospatial datasets:
tm_shape(mpsz_svy21) +
tm_polygons() +
tm_shape(condo_resale.sf) +
tm_dots()
2. Exploratory Data Analysis
Before building a model, we should examine the distribution of the variables in the data. The code below plots histograms for the numeric variables in the dataset.
To reduce repetitive code, we can create a function to plot histograms. It takes variable names as input. Since we are looping over a vector of strings of the variable names, we need to use the get() function to search for the corresponding object within the dataset. ggplot cannot accept string names.
autoplot <- function(x){
ggplot(data=condo_resale, aes(x= get(x))) +
geom_histogram(bins=20, color="black", fill="light blue")+
labs(x=x)
}
numdists <- map(names(condo_resale[4:20]), ~autoplot(.x))
ggarrange(plotlist=numdists, ncol=3, nrow=6)
Linear regression does not require independent variables to be normally distributed, but we should note if there are outliers. The dependent variables should also be linearly correlated with the indepdent variable.
We should also check for multicollienarity:
corrplot(cor(condo_resale[, 5:23]), diag = FALSE, order = "AOE",
tl.pos = "td", tl.cex = 0.5, method = "number", type = "upper")
LEASEHOLD_99YEAR has a strong negative correlation with FREEHOLD, so one of them should be excluded in the model. I will exclude LEASEHOLD_99YEAR in the subsequent model building.
By setting tmap_mode to view, we get an interactive map. We include the line tmap_options(check.and.fix = TRUE) because there are some invalid polygons. We can also use plot the values of a variable spatially by varying the colour of each point with respect to selling price. tm_dots() is used so that the size scales with the zoom.
tmap_mode("view")tmap mode set to interactive viewingtm_shape(mpsz_svy21)+
tmap_options(check.and.fix = TRUE)+
tm_polygons() +
tm_shape(condo_resale.sf) +
tm_dots(col = "SELLING_PRICE",
alpha = 0.6,
style="quantile") +
tm_view(set.zoom.limits = c(11,14))Warning: The shape mpsz_svy21 is invalid (after reprojection). See
sf::st_is_validSet the plot mode back to plot.
tmap_mode("plot")tmap mode set to plotting3. Regression Models
Simple Linear Regression
The following code creates a simple linear regression (only one explanatory variable). It models the relationship between selling price and size.
condo.slr <- lm(formula=SELLING_PRICE ~ AREA_SQM, data = condo_resale.sf)
summary(condo.slr)
Call:
lm(formula = SELLING_PRICE ~ AREA_SQM, data = condo_resale.sf)
Residuals:
Min 1Q Median 3Q Max
-3695815 -391764 -87517 258900 13503875
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -258121.1 63517.2 -4.064 5.09e-05 ***
AREA_SQM 14719.0 428.1 34.381 < 2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 942700 on 1434 degrees of freedom
Multiple R-squared: 0.4518, Adjusted R-squared: 0.4515
F-statistic: 1182 on 1 and 1434 DF, p-value: < 2.2e-16The adjusted R-squared value of 0.4515 means that the model can explain about 45% of the variation of the selling price. The p-value of the model is also low (<0.001) so we can reject the null hypothesis that mean selling price is a good estimator of selling price and accept the alternative hypothesis that the model is a better estimator of selling price.
Since we are only using 2 variables, we can easily plot the relationship:
ggplot(data=condo_resale.sf,
aes(x=`AREA_SQM`, y=`SELLING_PRICE`)) +
geom_point() +
geom_smooth(method = lm) +
theme_bw()`geom_smooth()` using formula = 'y ~ x'
While the results of the simple linear regression show that size can be used to estimate selling price, we may be able to estimate selling price better by including more explanatory variables.
Multiple Linear Regression
The following code chunk uses all the available independent variables:
condo.mlr <- lm(formula = SELLING_PRICE ~ AREA_SQM + AGE +
PROX_CBD + PROX_CHILDCARE + PROX_ELDERLYCARE +
PROX_URA_GROWTH_AREA + PROX_HAWKER_MARKET + PROX_KINDERGARTEN +
PROX_MRT + PROX_PARK + PROX_PRIMARY_SCH +
PROX_TOP_PRIMARY_SCH + PROX_SHOPPING_MALL + PROX_SUPERMARKET +
PROX_BUS_STOP + NO_Of_UNITS + FAMILY_FRIENDLY + FREEHOLD,
data=condo_resale.sf)
summary(condo.mlr)
Call:
lm(formula = SELLING_PRICE ~ AREA_SQM + AGE + PROX_CBD + PROX_CHILDCARE +
PROX_ELDERLYCARE + PROX_URA_GROWTH_AREA + PROX_HAWKER_MARKET +
PROX_KINDERGARTEN + PROX_MRT + PROX_PARK + PROX_PRIMARY_SCH +
PROX_TOP_PRIMARY_SCH + PROX_SHOPPING_MALL + PROX_SUPERMARKET +
PROX_BUS_STOP + NO_Of_UNITS + FAMILY_FRIENDLY + FREEHOLD,
data = condo_resale.sf)
Residuals:
Min 1Q Median 3Q Max
-3475964 -293923 -23069 241043 12260381
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 481728.40 121441.01 3.967 7.65e-05 ***
AREA_SQM 12708.32 369.59 34.385 < 2e-16 ***
AGE -24440.82 2763.16 -8.845 < 2e-16 ***
PROX_CBD -78669.78 6768.97 -11.622 < 2e-16 ***
PROX_CHILDCARE -351617.91 109467.25 -3.212 0.00135 **
PROX_ELDERLYCARE 171029.42 42110.51 4.061 5.14e-05 ***
PROX_URA_GROWTH_AREA 38474.53 12523.57 3.072 0.00217 **
PROX_HAWKER_MARKET 23746.10 29299.76 0.810 0.41782
PROX_KINDERGARTEN 147468.99 82668.87 1.784 0.07466 .
PROX_MRT -314599.68 57947.44 -5.429 6.66e-08 ***
PROX_PARK 563280.50 66551.68 8.464 < 2e-16 ***
PROX_PRIMARY_SCH 180186.08 65237.95 2.762 0.00582 **
PROX_TOP_PRIMARY_SCH 2280.04 20410.43 0.112 0.91107
PROX_SHOPPING_MALL -206604.06 42840.60 -4.823 1.57e-06 ***
PROX_SUPERMARKET -44991.80 77082.64 -0.584 0.55953
PROX_BUS_STOP 683121.35 138353.28 4.938 8.85e-07 ***
NO_Of_UNITS -231.18 89.03 -2.597 0.00951 **
FAMILY_FRIENDLY 140340.77 47020.55 2.985 0.00289 **
FREEHOLD 359913.01 49220.22 7.312 4.38e-13 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 755800 on 1417 degrees of freedom
Multiple R-squared: 0.6518, Adjusted R-squared: 0.6474
F-statistic: 147.4 on 18 and 1417 DF, p-value: < 2.2e-16The tbl_regression() function from the gtsummary package can create publication-ready summary tables for regression models. The following code chunk produces a clean coefficient table and the summary statistics of the model.
tbl_regression(condo.mlr,
intercept = TRUE) %>%
add_glance_source_note(
label = list(sigma ~ "\U03C3"),
include = c(r.squared, adj.r.squared,
AIC, statistic,
p.value, sigma))| Characteristic | Beta | 95% CI1 | p-value |
|---|---|---|---|
| (Intercept) | 481,728 | 243,505, 719,952 | <0.001 |
| AREA_SQM | 12,708 | 11,983, 13,433 | <0.001 |
| AGE | -24,441 | -29,861, -19,020 | <0.001 |
| PROX_CBD | -78,670 | -91,948, -65,391 | <0.001 |
| PROX_CHILDCARE | -351,618 | -566,353, -136,883 | 0.001 |
| PROX_ELDERLYCARE | 171,029 | 88,424, 253,635 | <0.001 |
| PROX_URA_GROWTH_AREA | 38,475 | 13,908, 63,041 | 0.002 |
| PROX_HAWKER_MARKET | 23,746 | -33,729, 81,222 | 0.4 |
| PROX_KINDERGARTEN | 147,469 | -14,698, 309,636 | 0.075 |
| PROX_MRT | -314,600 | -428,272, -200,928 | <0.001 |
| PROX_PARK | 563,280 | 432,730, 693,831 | <0.001 |
| PROX_PRIMARY_SCH | 180,186 | 52,213, 308,159 | 0.006 |
| PROX_TOP_PRIMARY_SCH | 2,280 | -37,758, 42,318 | >0.9 |
| PROX_SHOPPING_MALL | -206,604 | -290,642, -122,566 | <0.001 |
| PROX_SUPERMARKET | -44,992 | -196,200, 106,217 | 0.6 |
| PROX_BUS_STOP | 683,121 | 411,722, 954,521 | <0.001 |
| NO_Of_UNITS | -231 | -406, -57 | 0.010 |
| FAMILY_FRIENDLY | 140,341 | 48,103, 232,578 | 0.003 |
| FREEHOLD | 359,913 | 263,361, 456,465 | <0.001 |
| R² = 0.652; Adjusted R² = 0.647; AIC = 42,970; Statistic = 147; p-value = <0.001; σ = 755,816 | |||
| 1 CI = Confidence Interval | |||
The adjusted R-squared value of 0.647 means that the model can explain about 65% of the variation in selling price. This indicates that overall, this multiple linear regression model is a better estimator of the selling price than the simple linear regression model in the previous section.
However, we can also see that not all the variables used were statistically significant (alpha = 0.05), meaning that we cannot reject the null hypothesis that the coefficient for that variable is 0. Removal of these variables may improve fit.
condo.mlr1 <- lm(formula = SELLING_PRICE ~ AREA_SQM + AGE + PROX_CBD +
PROX_CHILDCARE + PROX_ELDERLYCARE +PROX_URA_GROWTH_AREA +
PROX_MRT + PROX_PARK + PROX_PRIMARY_SCH +
PROX_SHOPPING_MALL + PROX_BUS_STOP + NO_Of_UNITS +
FAMILY_FRIENDLY + FREEHOLD,
data=condo_resale.sf)
tbl_regression(condo.mlr1,
intercept = TRUE) %>%
add_glance_source_note(
label = list(sigma ~ "\U03C3"),
include = c(r.squared, adj.r.squared,
AIC, statistic,
p.value, sigma))| Characteristic | Beta | 95% CI1 | p-value |
|---|---|---|---|
| (Intercept) | 527,633 | 315,417, 739,849 | <0.001 |
| AREA_SQM | 12,778 | 12,057, 13,498 | <0.001 |
| AGE | -24,688 | -30,092, -19,284 | <0.001 |
| PROX_CBD | -77,131 | -88,436, -65,826 | <0.001 |
| PROX_CHILDCARE | -318,473 | -530,250, -106,696 | 0.003 |
| PROX_ELDERLYCARE | 185,576 | 107,303, 263,849 | <0.001 |
| PROX_URA_GROWTH_AREA | 39,163 | 16,105, 62,222 | <0.001 |
| PROX_MRT | -294,745 | -406,394, -183,096 | <0.001 |
| PROX_PARK | 570,505 | 442,004, 699,006 | <0.001 |
| PROX_PRIMARY_SCH | 159,856 | 41,698, 278,014 | 0.008 |
| PROX_SHOPPING_MALL | -220,947 | -292,668, -149,226 | <0.001 |
| PROX_BUS_STOP | 682,482 | 418,616, 946,348 | <0.001 |
| NO_Of_UNITS | -245 | -418, -73 | 0.005 |
| FAMILY_FRIENDLY | 146,308 | 54,321, 238,295 | 0.002 |
| FREEHOLD | 350,600 | 255,448, 445,752 | <0.001 |
| R² = 0.651; Adjusted R² = 0.647; AIC = 42,967; Statistic = 189; p-value = <0.001; σ = 755,957 | |||
| 1 CI = Confidence Interval | |||
Contrary to expectations, being nearer to parks and primary schools actually causes reduces the price of condos.
Check Regression Assumptions
There are a few main assumptions or requirements in linear regression:
No multicollinearity between explanatory variables
Linear relationship between independent and dependent variables
Independence of errors
Normality of errors
Equal variance of errors
The linearity assumption and independence of errors can be checked using the following code chunk. The points are approximately randomly scattered around the 0 line, indicating a linear relationship
ols_plot_resid_fit(condo.mlr1)
VIFs (variance inflation factor) can be used to check for multicollinearity. A VIF value >10 usually indicates that there is multicollinearity.
ols_vif_tol(condo.mlr1) Variables Tolerance VIF
1 AREA_SQM 0.8728554 1.145665
2 AGE 0.7071275 1.414172
3 PROX_CBD 0.6356147 1.573280
4 PROX_CHILDCARE 0.3066019 3.261559
5 PROX_ELDERLYCARE 0.6598479 1.515501
6 PROX_URA_GROWTH_AREA 0.7510311 1.331503
7 PROX_MRT 0.5236090 1.909822
8 PROX_PARK 0.8279261 1.207837
9 PROX_PRIMARY_SCH 0.4524628 2.210126
10 PROX_SHOPPING_MALL 0.6738795 1.483945
11 PROX_BUS_STOP 0.3514118 2.845664
12 NO_Of_UNITS 0.6901036 1.449058
13 FAMILY_FRIENDLY 0.7244157 1.380423
14 FREEHOLD 0.6931163 1.442759The following code chunk tests for unequal variance of errors (aka heteroskedasticity). With a low p-value, we reject the null hypothesis and conclude that the variance of errors is not constant.
ols_test_breusch_pagan(condo.mlr1)
Breusch Pagan Test for Heteroskedasticity
-----------------------------------------
Ho: the variance is constant
Ha: the variance is not constant
Data
-----------------------------------------
Response : SELLING_PRICE
Variables: fitted values of SELLING_PRICE
Test Summary
----------------------------
DF = 1
Chi2 = 4573.0707
Prob > Chi2 = 0.0000 Normality of errors can be tested by plotting the distribution of errors or using a statistical test. The low p-values mean that we reject the null hypothesis that the data (residuals) follow a normal distribution. In other words, the assumption for normality of residuals does not hold.
ols_plot_resid_hist(condo.mlr1)
ols_test_normality(condo.mlr1)Warning in ks.test.default(y, "pnorm", mean(y), sd(y)): ties should not be
present for the Kolmogorov-Smirnov test-----------------------------------------------
Test Statistic pvalue
-----------------------------------------------
Shapiro-Wilk 0.6856 0.0000
Kolmogorov-Smirnov 0.1366 0.0000
Cramer-von Mises 121.0768 0.0000
Anderson-Darling 67.9551 0.0000
-----------------------------------------------4. Spatial Autocorrelation
Geographically weighted regression estimates a dependent variable that varies across space. First we must test if the dependent variable does indeed exhibit systematic spatial variation, which is also know as spatial autocorrelation.
First, we need to convert the residuals from the multiple linear regression model from sf data frame into a spatial points data frame. The following code chunk extracts the residuals and appends it to the geosptial dataset.
mlr.output <- as.data.frame(condo.mlr1$residuals)
condo_resale.res.sf <- cbind(condo_resale.sf,
condo.mlr1$residuals) %>%
rename(`MLR_RES` = `condo.mlr1.residuals`)This code chunk converts it to a spatial points data frame object which is required by the GWmodel functions as input.
condo_resale.sp <- as_Spatial(condo_resale.res.sf)We can view the residuals on the map to visually look for spatial autocorrelation.
tmap_mode("view")tmap mode set to interactive viewingtm_shape(mpsz_svy21)+
tmap_options(check.and.fix = TRUE) +
tm_polygons(alpha = 0.4) +
tm_shape(condo_resale.res.sf) +
tm_dots(col = "MLR_RES",
alpha = 0.6,
style="quantile") +
tm_view(set.zoom.limits = c(11,14))Warning: The shape mpsz_svy21 is invalid (after reprojection). See
sf::st_is_validVariable(s) "MLR_RES" contains positive and negative values, so midpoint is set to 0. Set midpoint = NA to show the full spectrum of the color palette.There appears to be spatial autocorrelation. We can see that there are locations with clustering of observations with high residuals.
tmap_mode("plot")tmap mode set to plottingFor a more conclusive test, we can use the Moran’s I test to check if residuals are randomly distributed across space.
First, we need to define the neighbourhood. Since we are using point data, we will use a distance-based weight matrix. The following code chunk creates the list of neighbours of each point up to 1.5km away.
nb <- dnearneigh(coordinates(condo_resale.sp), 0, 1500, longlat = FALSE)
summary(nb)Neighbour list object:
Number of regions: 1436
Number of nonzero links: 66266
Percentage nonzero weights: 3.213526
Average number of links: 46.14624
Link number distribution:
1 3 5 7 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
3 3 9 4 3 15 10 19 17 45 19 5 14 29 19 6 35 45 18 47
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44
16 43 22 26 21 11 9 23 22 13 16 25 21 37 16 18 8 21 4 12
45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
8 36 18 14 14 43 11 12 8 13 12 13 4 5 6 12 11 20 29 33
65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84
15 20 10 14 15 15 11 16 12 10 8 19 12 14 9 8 4 13 11 6
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
4 9 4 4 4 6 2 16 9 4 5 9 3 9 4 2 1 2 1 1
105 106 107 108 109 110 112 116 125
1 5 9 2 1 3 1 1 1
3 least connected regions:
193 194 277 with 1 link
1 most connected region:
285 with 125 linksThen we convert it to a spatial weight matrix using nb2listw(). We are using a row-standardised weight matrix.
nb_lw <- nb2listw(nb, style = 'W')
summary(nb_lw)Characteristics of weights list object:
Neighbour list object:
Number of regions: 1436
Number of nonzero links: 66266
Percentage nonzero weights: 3.213526
Average number of links: 46.14624
Link number distribution:
1 3 5 7 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
3 3 9 4 3 15 10 19 17 45 19 5 14 29 19 6 35 45 18 47
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44
16 43 22 26 21 11 9 23 22 13 16 25 21 37 16 18 8 21 4 12
45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
8 36 18 14 14 43 11 12 8 13 12 13 4 5 6 12 11 20 29 33
65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84
15 20 10 14 15 15 11 16 12 10 8 19 12 14 9 8 4 13 11 6
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
4 9 4 4 4 6 2 16 9 4 5 9 3 9 4 2 1 2 1 1
105 106 107 108 109 110 112 116 125
1 5 9 2 1 3 1 1 1
3 least connected regions:
193 194 277 with 1 link
1 most connected region:
285 with 125 links
Weights style: W
Weights constants summary:
n nn S0 S1 S2
W 1436 2062096 1436 94.81916 5798.341Next, lm.morantest() of spdep package will be used to perform Moran’s I test for residual spatial autocorrelation.
lm.morantest(condo.mlr1, nb_lw)
Global Moran I for regression residuals
data:
model: lm(formula = SELLING_PRICE ~ AREA_SQM + AGE + PROX_CBD +
PROX_CHILDCARE + PROX_ELDERLYCARE + PROX_URA_GROWTH_AREA + PROX_MRT +
PROX_PARK + PROX_PRIMARY_SCH + PROX_SHOPPING_MALL + PROX_BUS_STOP +
NO_Of_UNITS + FAMILY_FRIENDLY + FREEHOLD, data = condo_resale.sf)
weights: nb_lw
Moran I statistic standard deviate = 24.366, p-value < 2.2e-16
alternative hypothesis: greater
sample estimates:
Observed Moran I Expectation Variance
1.438876e-01 -5.487594e-03 3.758259e-05 The p-value of the Moran’s I test is less than the alpha value of 0.05. We can reject the null hypothesis and conclude that residuals are not randomly distributed. Since the observed Moran’s I statistic is 0.14 which is >0, we can infer that there is clustering.
5. Building the GWR Model
Fixed Bandwith
The following code chunk is used to determine the optimal fixed bandwidth to use in the model. If we want to use adaptive distance (fixed number of neighbours), we would change the adaptive=FALSE argument to adaptive=TRUE. The approach argument sets the stopping rule to determine the maximum threshold distance.
bw.fixed <- bw.gwr(formula = SELLING_PRICE ~ AREA_SQM + AGE + PROX_CBD +
PROX_CHILDCARE + PROX_ELDERLYCARE + PROX_URA_GROWTH_AREA +
PROX_MRT + PROX_PARK + PROX_PRIMARY_SCH +
PROX_SHOPPING_MALL + PROX_BUS_STOP + NO_Of_UNITS +
FAMILY_FRIENDLY + FREEHOLD,
data=condo_resale.sp,
approach="CV",
kernel="gaussian",
adaptive=FALSE,
longlat=FALSE)Fixed bandwidth: 17660.96 CV score: 8.259118e+14
Fixed bandwidth: 10917.26 CV score: 7.970454e+14
Fixed bandwidth: 6749.419 CV score: 7.273273e+14
Fixed bandwidth: 4173.553 CV score: 6.300006e+14
Fixed bandwidth: 2581.58 CV score: 5.404958e+14
Fixed bandwidth: 1597.687 CV score: 4.857515e+14
Fixed bandwidth: 989.6077 CV score: 4.722431e+14
Fixed bandwidth: 613.7939 CV score: 1.378294e+16
Fixed bandwidth: 1221.873 CV score: 4.778717e+14
Fixed bandwidth: 846.0596 CV score: 4.791629e+14
Fixed bandwidth: 1078.325 CV score: 4.751406e+14
Fixed bandwidth: 934.7772 CV score: 4.72518e+14
Fixed bandwidth: 1023.495 CV score: 4.730305e+14
Fixed bandwidth: 968.6643 CV score: 4.721317e+14
Fixed bandwidth: 955.7206 CV score: 4.722072e+14
Fixed bandwidth: 976.6639 CV score: 4.721387e+14
Fixed bandwidth: 963.7202 CV score: 4.721484e+14
Fixed bandwidth: 971.7199 CV score: 4.721293e+14
Fixed bandwidth: 973.6083 CV score: 4.721309e+14
Fixed bandwidth: 970.5527 CV score: 4.721295e+14
Fixed bandwidth: 972.4412 CV score: 4.721296e+14
Fixed bandwidth: 971.2741 CV score: 4.721292e+14
Fixed bandwidth: 970.9985 CV score: 4.721293e+14
Fixed bandwidth: 971.4443 CV score: 4.721292e+14
Fixed bandwidth: 971.5496 CV score: 4.721293e+14
Fixed bandwidth: 971.3793 CV score: 4.721292e+14
Fixed bandwidth: 971.3391 CV score: 4.721292e+14
Fixed bandwidth: 971.3143 CV score: 4.721292e+14
Fixed bandwidth: 971.3545 CV score: 4.721292e+14
Fixed bandwidth: 971.3296 CV score: 4.721292e+14
Fixed bandwidth: 971.345 CV score: 4.721292e+14
Fixed bandwidth: 971.3355 CV score: 4.721292e+14
Fixed bandwidth: 971.3413 CV score: 4.721292e+14
Fixed bandwidth: 971.3377 CV score: 4.721292e+14
Fixed bandwidth: 971.34 CV score: 4.721292e+14
Fixed bandwidth: 971.3405 CV score: 4.721292e+14
Fixed bandwidth: 971.3408 CV score: 4.721292e+14
Fixed bandwidth: 971.3403 CV score: 4.721292e+14
Fixed bandwidth: 971.3406 CV score: 4.721292e+14
Fixed bandwidth: 971.3404 CV score: 4.721292e+14
Fixed bandwidth: 971.3405 CV score: 4.721292e+14
Fixed bandwidth: 971.3405 CV score: 4.721292e+14 The recommended bandwidth is 971.3405m.
The following code chunk performs the geographically weighted regression using the fixed bandwidth found in the previous step.
gwr.fixed <- gwr.basic(formula = SELLING_PRICE ~ AREA_SQM + AGE + PROX_CBD +
PROX_CHILDCARE + PROX_ELDERLYCARE + PROX_URA_GROWTH_AREA +
PROX_MRT + PROX_PARK + PROX_PRIMARY_SCH +
PROX_SHOPPING_MALL + PROX_BUS_STOP + NO_Of_UNITS +
FAMILY_FRIENDLY + FREEHOLD,
data=condo_resale.sp,
bw=bw.fixed,
kernel = 'gaussian',
longlat = FALSE)Let’s check the output.
gwr.fixed ***********************************************************************
* Package GWmodel *
***********************************************************************
Program starts at: 2022-12-10 15:09:19
Call:
gwr.basic(formula = SELLING_PRICE ~ AREA_SQM + AGE + PROX_CBD +
PROX_CHILDCARE + PROX_ELDERLYCARE + PROX_URA_GROWTH_AREA +
PROX_MRT + PROX_PARK + PROX_PRIMARY_SCH + PROX_SHOPPING_MALL +
PROX_BUS_STOP + NO_Of_UNITS + FAMILY_FRIENDLY + FREEHOLD,
data = condo_resale.sp, bw = bw.fixed, kernel = "gaussian",
longlat = FALSE)
Dependent (y) variable: SELLING_PRICE
Independent variables: AREA_SQM AGE PROX_CBD PROX_CHILDCARE PROX_ELDERLYCARE PROX_URA_GROWTH_AREA PROX_MRT PROX_PARK PROX_PRIMARY_SCH PROX_SHOPPING_MALL PROX_BUS_STOP NO_Of_UNITS FAMILY_FRIENDLY FREEHOLD
Number of data points: 1436
***********************************************************************
* Results of Global Regression *
***********************************************************************
Call:
lm(formula = formula, data = data)
Residuals:
Min 1Q Median 3Q Max
-3470778 -298119 -23481 248917 12234210
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 527633.22 108183.22 4.877 1.20e-06 ***
AREA_SQM 12777.52 367.48 34.771 < 2e-16 ***
AGE -24687.74 2754.84 -8.962 < 2e-16 ***
PROX_CBD -77131.32 5763.12 -13.384 < 2e-16 ***
PROX_CHILDCARE -318472.75 107959.51 -2.950 0.003231 **
PROX_ELDERLYCARE 185575.62 39901.86 4.651 3.61e-06 ***
PROX_URA_GROWTH_AREA 39163.25 11754.83 3.332 0.000885 ***
PROX_MRT -294745.11 56916.37 -5.179 2.56e-07 ***
PROX_PARK 570504.81 65507.03 8.709 < 2e-16 ***
PROX_PRIMARY_SCH 159856.14 60234.60 2.654 0.008046 **
PROX_SHOPPING_MALL -220947.25 36561.83 -6.043 1.93e-09 ***
PROX_BUS_STOP 682482.22 134513.24 5.074 4.42e-07 ***
NO_Of_UNITS -245.48 87.95 -2.791 0.005321 **
FAMILY_FRIENDLY 146307.58 46893.02 3.120 0.001845 **
FREEHOLD 350599.81 48506.48 7.228 7.98e-13 ***
---Significance stars
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 756000 on 1421 degrees of freedom
Multiple R-squared: 0.6507
Adjusted R-squared: 0.6472
F-statistic: 189.1 on 14 and 1421 DF, p-value: < 2.2e-16
***Extra Diagnostic information
Residual sum of squares: 8.120609e+14
Sigma(hat): 752522.9
AIC: 42966.76
AICc: 42967.14
BIC: 41731.39
***********************************************************************
* Results of Geographically Weighted Regression *
***********************************************************************
*********************Model calibration information*********************
Kernel function: gaussian
Fixed bandwidth: 971.3405
Regression points: the same locations as observations are used.
Distance metric: Euclidean distance metric is used.
****************Summary of GWR coefficient estimates:******************
Min. 1st Qu. Median 3rd Qu.
Intercept -3.5988e+07 -5.1998e+05 7.6780e+05 1.7412e+06
AREA_SQM 1.0003e+03 5.2758e+03 7.4740e+03 1.2301e+04
AGE -1.3475e+05 -2.0813e+04 -8.6260e+03 -3.7784e+03
PROX_CBD -7.7047e+07 -2.3608e+05 -8.3600e+04 3.4646e+04
PROX_CHILDCARE -6.0097e+06 -3.3667e+05 -9.7425e+04 2.9007e+05
PROX_ELDERLYCARE -3.5000e+06 -1.5970e+05 3.1971e+04 1.9577e+05
PROX_URA_GROWTH_AREA -3.0170e+06 -8.2013e+04 7.0749e+04 2.2612e+05
PROX_MRT -3.5282e+06 -6.5836e+05 -1.8833e+05 3.6922e+04
PROX_PARK -1.2062e+06 -2.1732e+05 3.5383e+04 4.1335e+05
PROX_PRIMARY_SCH -2.2695e+07 -1.7066e+05 4.8472e+04 5.1555e+05
PROX_SHOPPING_MALL -7.2585e+06 -1.6684e+05 -1.0517e+04 1.5923e+05
PROX_BUS_STOP -1.4676e+06 -4.5207e+04 3.7601e+05 1.1664e+06
NO_Of_UNITS -1.3170e+03 -2.4822e+02 -3.0846e+01 2.5496e+02
FAMILY_FRIENDLY -2.2749e+06 -1.1140e+05 7.6214e+03 1.6107e+05
FREEHOLD -9.2067e+06 3.8073e+04 1.5169e+05 3.7528e+05
Max.
Intercept 112793548
AREA_SQM 21575
AGE 434201
PROX_CBD 2704596
PROX_CHILDCARE 1654087
PROX_ELDERLYCARE 38867814
PROX_URA_GROWTH_AREA 78515730
PROX_MRT 3124316
PROX_PARK 18122425
PROX_PRIMARY_SCH 4637503
PROX_SHOPPING_MALL 1529952
PROX_BUS_STOP 11342182
NO_Of_UNITS 12907
FAMILY_FRIENDLY 1720744
FREEHOLD 6073636
************************Diagnostic information*************************
Number of data points: 1436
Effective number of parameters (2trace(S) - trace(S'S)): 438.3804
Effective degrees of freedom (n-2trace(S) + trace(S'S)): 997.6196
AICc (GWR book, Fotheringham, et al. 2002, p. 61, eq 2.33): 42263.61
AIC (GWR book, Fotheringham, et al. 2002,GWR p. 96, eq. 4.22): 41632.36
BIC (GWR book, Fotheringham, et al. 2002,GWR p. 61, eq. 2.34): 42515.71
Residual sum of squares: 2.53407e+14
R-square value: 0.8909912
Adjusted R-square value: 0.8430417
***********************************************************************
Program stops at: 2022-12-10 15:09:20 The AIC of the GWR model is 41632.36 , which is lower than the AIC of the non-GWR model of 42966.76. This means that the GWR is a better model. The adjusted R-squared of the model is 0.8430417 which is higher than the R-square of the multiple linear regression model’s R-square of 0.647. This means that the geographically weighted model is able to explain more of the variation in selling price than the multiple linear regression model.
Adaptive Bandwidth
The following code chunk uses adaptive bandwidth instead. It recommends the number of points that should be considered as neighbours.
bw.adaptive <- bw.gwr(formula = SELLING_PRICE ~ AREA_SQM + AGE +
PROX_CBD + PROX_CHILDCARE + PROX_ELDERLYCARE +
PROX_URA_GROWTH_AREA + PROX_MRT + PROX_PARK +
PROX_PRIMARY_SCH + PROX_SHOPPING_MALL + PROX_BUS_STOP +
NO_Of_UNITS + FAMILY_FRIENDLY + FREEHOLD,
data=condo_resale.sp,
approach="CV",
kernel="gaussian",
adaptive=TRUE,
longlat=FALSE)Adaptive bandwidth: 895 CV score: 7.952401e+14
Adaptive bandwidth: 561 CV score: 7.667364e+14
Adaptive bandwidth: 354 CV score: 6.953454e+14
Adaptive bandwidth: 226 CV score: 6.15223e+14
Adaptive bandwidth: 147 CV score: 5.674373e+14
Adaptive bandwidth: 98 CV score: 5.426745e+14
Adaptive bandwidth: 68 CV score: 5.168117e+14
Adaptive bandwidth: 49 CV score: 4.859631e+14
Adaptive bandwidth: 37 CV score: 4.646518e+14
Adaptive bandwidth: 30 CV score: 4.422088e+14
Adaptive bandwidth: 25 CV score: 4.430816e+14
Adaptive bandwidth: 32 CV score: 4.505602e+14
Adaptive bandwidth: 27 CV score: 4.462172e+14
Adaptive bandwidth: 30 CV score: 4.422088e+14 The algorithm recommends 30 as the optimal number of neighbours.
Now, we can conduct the geographically weighted regression.
gwr.adaptive <- gwr.basic(formula = SELLING_PRICE ~ AREA_SQM + AGE +
PROX_CBD + PROX_CHILDCARE + PROX_ELDERLYCARE +
PROX_URA_GROWTH_AREA + PROX_MRT + PROX_PARK +
PROX_PRIMARY_SCH + PROX_SHOPPING_MALL + PROX_BUS_STOP +
NO_Of_UNITS + FAMILY_FRIENDLY + FREEHOLD,
data=condo_resale.sp,
bw=bw.adaptive,
kernel = 'gaussian',
adaptive=TRUE,
longlat = FALSE)gwr.adaptive ***********************************************************************
* Package GWmodel *
***********************************************************************
Program starts at: 2022-12-10 15:09:26
Call:
gwr.basic(formula = SELLING_PRICE ~ AREA_SQM + AGE + PROX_CBD +
PROX_CHILDCARE + PROX_ELDERLYCARE + PROX_URA_GROWTH_AREA +
PROX_MRT + PROX_PARK + PROX_PRIMARY_SCH + PROX_SHOPPING_MALL +
PROX_BUS_STOP + NO_Of_UNITS + FAMILY_FRIENDLY + FREEHOLD,
data = condo_resale.sp, bw = bw.adaptive, kernel = "gaussian",
adaptive = TRUE, longlat = FALSE)
Dependent (y) variable: SELLING_PRICE
Independent variables: AREA_SQM AGE PROX_CBD PROX_CHILDCARE PROX_ELDERLYCARE PROX_URA_GROWTH_AREA PROX_MRT PROX_PARK PROX_PRIMARY_SCH PROX_SHOPPING_MALL PROX_BUS_STOP NO_Of_UNITS FAMILY_FRIENDLY FREEHOLD
Number of data points: 1436
***********************************************************************
* Results of Global Regression *
***********************************************************************
Call:
lm(formula = formula, data = data)
Residuals:
Min 1Q Median 3Q Max
-3470778 -298119 -23481 248917 12234210
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 527633.22 108183.22 4.877 1.20e-06 ***
AREA_SQM 12777.52 367.48 34.771 < 2e-16 ***
AGE -24687.74 2754.84 -8.962 < 2e-16 ***
PROX_CBD -77131.32 5763.12 -13.384 < 2e-16 ***
PROX_CHILDCARE -318472.75 107959.51 -2.950 0.003231 **
PROX_ELDERLYCARE 185575.62 39901.86 4.651 3.61e-06 ***
PROX_URA_GROWTH_AREA 39163.25 11754.83 3.332 0.000885 ***
PROX_MRT -294745.11 56916.37 -5.179 2.56e-07 ***
PROX_PARK 570504.81 65507.03 8.709 < 2e-16 ***
PROX_PRIMARY_SCH 159856.14 60234.60 2.654 0.008046 **
PROX_SHOPPING_MALL -220947.25 36561.83 -6.043 1.93e-09 ***
PROX_BUS_STOP 682482.22 134513.24 5.074 4.42e-07 ***
NO_Of_UNITS -245.48 87.95 -2.791 0.005321 **
FAMILY_FRIENDLY 146307.58 46893.02 3.120 0.001845 **
FREEHOLD 350599.81 48506.48 7.228 7.98e-13 ***
---Significance stars
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 756000 on 1421 degrees of freedom
Multiple R-squared: 0.6507
Adjusted R-squared: 0.6472
F-statistic: 189.1 on 14 and 1421 DF, p-value: < 2.2e-16
***Extra Diagnostic information
Residual sum of squares: 8.120609e+14
Sigma(hat): 752522.9
AIC: 42966.76
AICc: 42967.14
BIC: 41731.39
***********************************************************************
* Results of Geographically Weighted Regression *
***********************************************************************
*********************Model calibration information*********************
Kernel function: gaussian
Adaptive bandwidth: 30 (number of nearest neighbours)
Regression points: the same locations as observations are used.
Distance metric: Euclidean distance metric is used.
****************Summary of GWR coefficient estimates:******************
Min. 1st Qu. Median 3rd Qu.
Intercept -1.3487e+08 -2.4669e+05 7.7928e+05 1.6194e+06
AREA_SQM 3.3188e+03 5.6285e+03 7.7825e+03 1.2738e+04
AGE -9.6746e+04 -2.9288e+04 -1.4043e+04 -5.6119e+03
PROX_CBD -2.5330e+06 -1.6256e+05 -7.7242e+04 2.6624e+03
PROX_CHILDCARE -1.2790e+06 -2.0175e+05 8.7158e+03 3.7778e+05
PROX_ELDERLYCARE -1.6212e+06 -9.2050e+04 6.1029e+04 2.8184e+05
PROX_URA_GROWTH_AREA -7.2686e+06 -3.0350e+04 4.5869e+04 2.4613e+05
PROX_MRT -4.3781e+07 -6.7282e+05 -2.2115e+05 -7.4593e+04
PROX_PARK -2.9020e+06 -1.6782e+05 1.1601e+05 4.6572e+05
PROX_PRIMARY_SCH -8.6418e+05 -1.6627e+05 -7.7853e+03 4.3222e+05
PROX_SHOPPING_MALL -1.8272e+06 -1.3175e+05 -1.4049e+04 1.3799e+05
PROX_BUS_STOP -2.0579e+06 -7.1461e+04 4.1104e+05 1.2071e+06
NO_Of_UNITS -2.1993e+03 -2.3685e+02 -3.4699e+01 1.1657e+02
FAMILY_FRIENDLY -5.9879e+05 -5.0927e+04 2.6173e+04 2.2481e+05
FREEHOLD -1.6340e+05 4.0765e+04 1.9023e+05 3.7960e+05
Max.
Intercept 18758355
AREA_SQM 23064
AGE 13303
PROX_CBD 11346650
PROX_CHILDCARE 2892127
PROX_ELDERLYCARE 2465671
PROX_URA_GROWTH_AREA 7384059
PROX_MRT 1186242
PROX_PARK 2588497
PROX_PRIMARY_SCH 3381462
PROX_SHOPPING_MALL 38038564
PROX_BUS_STOP 12081592
NO_Of_UNITS 1010
FAMILY_FRIENDLY 2072414
FREEHOLD 1813995
************************Diagnostic information*************************
Number of data points: 1436
Effective number of parameters (2trace(S) - trace(S'S)): 350.3088
Effective degrees of freedom (n-2trace(S) + trace(S'S)): 1085.691
AICc (GWR book, Fotheringham, et al. 2002, p. 61, eq 2.33): 41982.22
AIC (GWR book, Fotheringham, et al. 2002,GWR p. 96, eq. 4.22): 41546.74
BIC (GWR book, Fotheringham, et al. 2002,GWR p. 61, eq. 2.34): 41914.08
Residual sum of squares: 2.528227e+14
R-square value: 0.8912425
Adjusted R-square value: 0.8561185
***********************************************************************
Program stops at: 2022-12-10 15:09:27 AIC of the GWR model is also lower than that of the non-GWR model, so it is a better model. The R-square of the GWR model using adaptive distance is 0.856 which is also better than that of the MLR model (0.647), meaning that the GWR model can explain selling price better.
To compare between the 2 GWR models, we should compare AIC instead. The AIC of the adaptive distance model is 41546.74 and the AIC of the fixed distance model is 41632.36 . The adaptive distance model is slightly better.
Visualising GWR Output
On top of the printed output shown in the previous sections, the gwr.basic() function also outputs some other diagnostics saved as a spatial points data frame named SDF.
Condition Number: this diagnostic evaluates local collinearity. In the presence of strong local collinearity, results become unstable. Results associated with condition numbers larger than 30, may be unreliable.
Local R2: these values range between 0.0 and 1.0 and indicate how well the local regression model fits observed y values. Very low values indicate the local model is performing poorly. Mapping the Local R2 values to see where GWR predicts well and where it predicts poorly may provide clues about important variables that may be missing from the regression model.
Predicted y (yhat): these are the estimated (or fitted) y values computed by GWR.
Residuals: Residuals are the difference between observed y values and the fitted y values. Standardized residuals have a mean of zero and a standard deviation of 1. A cold-to-hot rendered map of standardized residuals can be produce by using these values.
Coefficient Standard Error: these values measure the reliability of each coefficient estimate. Confidence in those estimates are higher when standard errors are small in relation to the actual coefficient values. Large standard errors may indicate problems with local collinearity.
To visualise the results of the GWR model, we extract it and convert it to an sf object.
condo_resale.sf.adaptive <- st_as_sf(gwr.adaptive$SDF) %>%
st_transform(crs=3414)glimpse(condo_resale.sf.adaptive)Rows: 1,436
Columns: 52
$ Intercept <dbl> 2050011.67, 1633128.24, 3433608.17, 234358.91,…
$ AREA_SQM <dbl> 9561.892, 16576.853, 13091.861, 20730.601, 672…
$ AGE <dbl> -9514.634, -58185.479, -26707.386, -93308.988,…
$ PROX_CBD <dbl> -120681.94, -149434.22, -259397.77, 2426853.66…
$ PROX_CHILDCARE <dbl> 319266.925, 441102.177, -120116.816, 480825.28…
$ PROX_ELDERLYCARE <dbl> -393417.795, 325188.741, 535855.806, 314783.72…
$ PROX_URA_GROWTH_AREA <dbl> -159980.203, -142290.389, -253621.206, -267929…
$ PROX_MRT <dbl> -299742.96, -2510522.23, -936853.28, -2039479.…
$ PROX_PARK <dbl> -172104.47, 523379.72, 209099.85, -759153.26, …
$ PROX_PRIMARY_SCH <dbl> 242668.03, 1106830.66, 571462.33, 3127477.21, …
$ PROX_SHOPPING_MALL <dbl> 300881.390, -87693.378, -126732.712, -29593.34…
$ PROX_BUS_STOP <dbl> 1210615.44, 1843587.22, 1411924.90, 7225577.51…
$ NO_Of_UNITS <dbl> 104.8290640, -288.3441183, -9.5532945, -161.35…
$ FAMILY_FRIENDLY <dbl> -9075.370, 310074.664, 5949.746, 1556178.531, …
$ FREEHOLD <dbl> 303955.61, 396221.27, 168821.75, 1212515.58, 3…
$ y <dbl> 3000000, 3880000, 3325000, 4250000, 1400000, 1…
$ yhat <dbl> 2886531.8, 3466801.5, 3616527.2, 5435481.6, 13…
$ residual <dbl> 113468.16, 413198.52, -291527.20, -1185481.63,…
$ CV_Score <dbl> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ Stud_residual <dbl> 0.38207013, 1.01433140, -0.83780678, -2.846146…
$ Intercept_SE <dbl> 516105.5, 488083.5, 963711.4, 444185.5, 211962…
$ AREA_SQM_SE <dbl> 823.2860, 825.2380, 988.2240, 617.4007, 1376.2…
$ AGE_SE <dbl> 5889.782, 6226.916, 6510.236, 6010.511, 8180.3…
$ PROX_CBD_SE <dbl> 37411.22, 23615.06, 56103.77, 469337.41, 41064…
$ PROX_CHILDCARE_SE <dbl> 319111.1, 299705.3, 349128.5, 304965.2, 698720…
$ PROX_ELDERLYCARE_SE <dbl> 120633.34, 84546.69, 129687.07, 127150.69, 327…
$ PROX_URA_GROWTH_AREA_SE <dbl> 56207.39, 76956.50, 95774.60, 470762.12, 47433…
$ PROX_MRT_SE <dbl> 185181.3, 281133.9, 275483.7, 279877.1, 363830…
$ PROX_PARK_SE <dbl> 205499.6, 229358.7, 314124.3, 227249.4, 364580…
$ PROX_PRIMARY_SCH_SE <dbl> 152400.7, 165150.7, 196662.6, 240878.9, 249087…
$ PROX_SHOPPING_MALL_SE <dbl> 109268.8, 98906.8, 119913.3, 177104.1, 301032.…
$ PROX_BUS_STOP_SE <dbl> 600668.6, 410222.1, 464156.7, 562810.8, 740922…
$ NO_Of_UNITS_SE <dbl> 218.1258, 208.9410, 210.9828, 361.7767, 299.50…
$ FAMILY_FRIENDLY_SE <dbl> 131474.73, 114989.07, 146607.22, 108726.62, 16…
$ FREEHOLD_SE <dbl> 115954.0, 130110.0, 141031.5, 138239.1, 210641…
$ Intercept_TV <dbl> 3.9720784, 3.3460017, 3.5629010, 0.5276150, 1.…
$ AREA_SQM_TV <dbl> 11.614302, 20.087361, 13.247868, 33.577223, 4.…
$ AGE_TV <dbl> -1.6154474, -9.3441881, -4.1023685, -15.524301…
$ PROX_CBD_TV <dbl> -3.22582173, -6.32792021, -4.62353528, 5.17080…
$ PROX_CHILDCARE_TV <dbl> 1.000488185, 1.471786337, -0.344047555, 1.5766…
$ PROX_ELDERLYCARE_TV <dbl> -3.26126929, 3.84626245, 4.13191383, 2.4756745…
$ PROX_URA_GROWTH_AREA_TV <dbl> -2.846248368, -1.848971738, -2.648105057, -5.6…
$ PROX_MRT_TV <dbl> -1.61864578, -8.92998600, -3.40075727, -7.2870…
$ PROX_PARK_TV <dbl> -0.83749312, 2.28192684, 0.66565951, -3.340617…
$ PROX_PRIMARY_SCH_TV <dbl> 1.59230221, 6.70194543, 2.90580089, 12.9836104…
$ PROX_SHOPPING_MALL_TV <dbl> 2.753588422, -0.886626400, -1.056869486, -0.16…
$ PROX_BUS_STOP_TV <dbl> 2.0154464, 4.4941192, 3.0419145, 12.8383775, 0…
$ NO_Of_UNITS_TV <dbl> 0.480589953, -1.380026395, -0.045279967, -0.44…
$ FAMILY_FRIENDLY_TV <dbl> -0.06902748, 2.69655779, 0.04058290, 14.312764…
$ FREEHOLD_TV <dbl> 2.6213469, 3.0452799, 1.1970499, 8.7711485, 1.…
$ Local_R2 <dbl> 0.8846744, 0.8899773, 0.8947007, 0.9073605, 0.…
$ geometry <POINT [m]> POINT (22085.12 29951.54), POINT (25656.…The following code chunk produces a summary of the predicted y values.
summary(gwr.adaptive$SDF$yhat) Min. 1st Qu. Median Mean 3rd Qu. Max.
171347 1102001 1385528 1751842 1982307 13887901 The following code chunk visualises local R-square of each observation.
tmap_mode("view")tmap mode set to interactive viewingtm_shape(mpsz_svy21)+
tm_polygons(alpha = 0.1) +
tm_shape(condo_resale.sf.adaptive) +
tm_dots(col = "Local_R2",
border.col = "gray60",
border.lwd = 1) +
tm_view(set.zoom.limits = c(11,14))Warning: The shape mpsz_svy21 is invalid (after reprojection). See
sf::st_is_validWe can see that the GWR model is better at predicting selling prices in some areas than others. Specifically, the model is not as accurate at predicting prices near the centre of Singapore. There could be that there is an unseen variable that is not included in the model that could explain this, or that we should build a different model for the region.
We can take a closer look by restricting the planning areas. These values are specifically in the Bishan, Toa Payoh, Novena and Kallang. Note that there are still points outside these regions as we did not filter the results for the condo_resale.sf.adaptive dataset. It would need to be related to the subzone data to conduct the filtering.
tmap_mode("plot")tmap mode set to plottingtm_shape(mpsz_svy21[mpsz_svy21$PLN_AREA_C %in% c("BS", "TP", "NV", "KL"), ])+
tm_polygons()+
tm_shape(condo_resale.sf.adaptive) +
tm_bubbles(col = "Local_R2",
size = 0.15,
border.col = "gray60",
border.lwd = 1)
It is also useful to visualise the local coefficient estimates. The gwr.basic() function stores the coefficient estimates, standard error and t-values with no suffix, “_SE” suffix and “_TV” suffix respectively. The following code chunk uses the pt() function to find the p-values from the t-values to determine which estimates are statistically significant. We want to know if the coefficient is different from zero, so we use a 2-tailed t-test. To do so, we use the absolute t-value and multiply the p-value by 2.
condo_resale.sf.adaptive <- condo_resale.sf.adaptive %>%
mutate(across(AREA_SQM_TV:FREEHOLD_TV,
~pt(abs(.x), df=gwr.adaptive$GW.diagnostic$edf,
lower.tail=FALSE)*2,
.names = "{col}_p")) %>%
rename_with(~gsub("_TV_","_", .x, fixed=TRUE))The following code chunk plot plots the local coefficient estimate and the corresponding p-values for AREA_SQM only. We can plot the maps for any of the variables used the values were computed in the last step.
tmap_mode("view")tmap mode set to interactive viewingAREA_SQM_SE <- tm_shape(mpsz_svy21)+
tm_polygons(alpha = 0.1) +
tm_shape(condo_resale.sf.adaptive) +
tm_dots(col = "AREA_SQM",
border.col = "gray60",
border.lwd = 1) +
tm_view(set.zoom.limits = c(11,14))
AREA_SQM_TV <- tm_shape(mpsz_svy21)+
tm_polygons(alpha = 0.1) +
tm_shape(condo_resale.sf.adaptive) +
tm_dots(col = "AREA_SQM_p",
style= "fixed",
breaks= c(0, 0.001, 0.01, 0.05, 1),
palette = "-Oranges",
border.col = "gray60",
border.lwd = 1) +
tm_view(set.zoom.limits = c(11,14))
tmap_arrange(AREA_SQM_SE, AREA_SQM_TV,
asp=1, ncol=2,
sync = TRUE)Warning: The shape mpsz_svy21 is invalid (after reprojection). See
sf::st_is_valid
Warning: The shape mpsz_svy21 is invalid (after reprojection). See
sf::st_is_valid